1 A Comparison of Approaches to Large-Scale Data Analysis Pavlo, Paulson, Rasin, Abadi, DeWitt, Madden, Stonebraker, SIGMOD’09 Shimin Chen Big data reading group

2 Two Approaches to Large-Scale Data Analysis “Shared nothing” MapReduce  Distributed file system  Map, Split, Copy, Reduce  MR scheduler Parallel DBMS  Standard relational tables, (physical location transparent)  Data are partitioned over cluster nodes  SQL  Join processing: T1 joins T2 If T2 is small, copy T2 to all the machines If T2 is large, then hash partition T1 and T2 and send partitions to different machines (this is similar to the split-copy in MapReduce)

3 Comparing the Two Approaches Schema Support Programming Model & Flexibility Indexing Execution Strategy & Fault Tolerance Data transfers

4 Schema Support MapReduce  Flexible, programmers write code to interpret input data  Good for single application scenario  Bad if data are shared by multiple applications. Must address data syntax, consistency, etc. Parallel DBMS  Relational schema required  Good if data are shared by multiple applications

5 Programming Model & Flexibility MapReduce  Low level  “Anecdotal evidence from the MR community suggests that there is widespread sharing of MR code fragments to do common tasks, such as joining data sets.”  very flexible Parallel DBMS  SQL  user-defined functions, stored procedures, user-defined aggregates

6 Indexing MapReduce  No native index support  Programmers can implement their own index support in Map/Reduce code  But hard to share the customized indexes in multiple applications Parallel DBMS  Hash/b-tree indexes well supported

7 Execution Strategy & Fault Tolerance MapReduce  Intermediate results are saved to local files  If a node fails, run the node-task again on another node  At a mapper machine, when multiple reducers are reading multiple local files, there could be large numbers of disk seeks, leading to poor performance. Parallel DBMS  Intermediate results are pushed across network  If a node fails, must re-run the entire query

8 Avoiding Data Transfers MapReduce  Schedule Map to close to data  But other than this, programmers must avoid data transfers themselves Parallel DBMS  A lot of optimizations  Such as determine where to perform filtering

9 Performance Benchmarks Benchmark Environment Original MR task (Grep) Analytical Tasks  Selection  Aggregation  Join  User-defined-function (UDF) aggregation

10 Node Configuration 100-node cluster Each node: 2.40GHz Intel Core 2 Duo, 64-bit red hat enterprise linux 5 (kernel ) w/ 4Gb RAM and two 250GB SATA HDDs. Nodes are connected to Cisco Catalyst 3750E 1Gbps switches. Internal switching fabric has 128Gbps. 50 nodes per switch. Multiple switches are connected to create a 64Gbps Cisco StackWise plus ring. The ring is only used for cross-switch communications.

11 Tested Systems Hadoop ( on Java 1.6.0)  HDFS data block size: 256MB  JVMs use 3.5GB heap size per node  “Rack awareness” enabled for data locality  Three replicas w/o compression: Compression or fewer replicas in HDFS does not improve performance DBMS-X (a parallel SQL DBMS from a major vendor)  Row store  4GB shared memory for buffer pool and temp space per node  Compressed table (compression often reduces time by 50%) Vertica  Column store  256MB buffer size per node  Compressed columns by default

12 Benchmark Execution Data loading time:  Actual loading of the data  Additional operations after the loading, such as compressing or building indexes Execution time  DBMS-X and vertica: Final results are piped from a shell command into a file  Hadoop: Final results are stored in HDFS An additional step to combine the multiple files into a single file

13 Performance Benchmarks Benchmark Environment Original MR task (Grep) Analytical Tasks  Selection  Aggregation  Join  User-defined-function (UDF) aggregation

14 Task Description From MapReduce paper  Input data set: 100-byte records  Look for a three-character pattern  One match per 10,000 records Varying the number of nodes  Fix the size of data per node (535MB/node)  Fix the total data size (1TB)

15 Data Loading Hadoop:  Copy text files into HDFS in parallel DBMS-X:  Load SQL command executed in parallel: it performs hash partitioning and distributes records to multiple machines  Reorganize data on each node: compress data, build index, perform other housekeeping This happens in parallel Vertica:  Copy command to load data in parallel  Data is re-distributed, then compressed

16 Data Loading Times DBMS-X: grey is loading, white is re-organization after loading Loading is actually sequential despite parallel load commands Hadoop does better because it only copies the data to three HDFS replicas.

17 Execution SQL:  SELECT * FROM data WHERE field LIKE “%XYZ%”  Full table scan MapReduce:  Map: pattern search  No reduce  An additional MR job to combine the output into a single file

18 Execution time Hadoop’s large start-up cost shows up in Figure 4, when data per node is small Vertica’s good data compression Combine output grep

19 Performance Benchmarks Benchmark Environment Original MR task (Grep) Analytical Tasks  Selection  Aggregation  Join  User-defined-function (UDF) aggregation

20 Input Data Input #1: random HTML documents  Inside an html doc, links are generated with Zipfian distribution  600,000 unique html docs with unique urls per node Input #2: 155 million UserVisits records  20GB/node Input #3: 18 million Ranking records  1GB/node

21 Selection Task Find the pageURLs in the rankings table (1GB/node) with a pageRank > threshold  36,000 records per data file (very selective) SQL: SELECT pageURL, pageRank FROM Rankings WHERE pageRank > X; MR: single Map, no Reduce

22 Hadoop’s start-up cost; DBMS uses index; vertica’s reliable message layer becomes bottleneck

23 Aggregation Task Calculate the total adRevenue generated for each sourceIP in the UserVisits table (20GB/node), grouped by the sourceIP column.  Nodes must exchange info for computing groupby  Generate 53 MB data regardless of number of nodes SQL: SELECT sourceIP, SUM(adRevenue) FROM UserVisits GROUP BY sourceIP; MR:  Map: outputs (sourceIP, adRevenue)  Reduce: compute sum per sourceIP  “Combine” is used

24 DBMS: Local group-by, then the coordinator performs the global group-by; performance dominated by data transfer.

25 Join Task Find the sourceIP that generated the most revenue within Jan 15-22, 2000, then calculate the average pageRank of all the pages visited by the sourceIP during this interval. SQL: SELECT INTO Temp sourceIP, AVG(pageRank) as avgPageRank, SUM(adRevenue) as totalRevenue FROM Rankings AS R, UserVisits AS UV WHERE R.pageURL = UV.destURL AND UV.visitDate BETWEEN Date(‘ ’) AND Date(‘ ’) GROUP BY UV.sourceIP; SELECT sourceIP, totalRevenue, avgPageRank FROM Temp ORDER BY totalRevenue DESC LIMIT 1;

26 MR Phase 1: filter UserVisits that are outside the desired date range, joins the qualifying records with records from the Ranking file. Phase 2: compute total adRevenue and average pageRank per sourceIP Phase 3: produce the largest record

27 DBMS can use index, both relations are partitioned on the join key; MR has to read all data. MR phase 1 takes an average seconds 600 seconds of raw I/O to read the table; 300 seconds to split, parse, deserialize; Thus CPU overhead is the limiting factor

28 UDF Aggregation Task Compute inlink count per document SQL: SELECT INTO Temp F(contents) FROM Documents; SELECT url, SUM(value) FROM Temp GROUP BY url; Need a user-defined-function to parse an html doc. Both DBMS’s do not support UDF very well. For example, get around by opening the html doc from local disk rather than perform all processing in DBMS. MR:  A standard MR program

29 DBMS: lower – UDF time; upper – other query time Hadoop: lower – query time; upper: combine all results into one

30 Discussion Parallel DBMSs much more challenging than Hadoop to install and configure properly Hadoop task start-up cost  100 node system: 10 second till the first task starts, 25 seconds till all nodes run tasks Compression: does not help in Hadoop? Execution strategy, failure model, ease of use